CN101237469A - Method for optimizing multi-QoS grid workflow based on ant group algorithm - Google Patents

Abstract

The invention discloses a method for optimizing a plurality of QoS grid workflows by adopting an ACO, wherein, three kinds of QoS parameters are considered. The method can not only satisfy the QoS constraint given by a user but also optimize one QoS parameter given by the user. The method comprises the following steps that: firstly, ants are initialized, each ant can select a structural direction in random and establishes a task sequence in accordance with the preference constraint; next, each ant selects a kind of heuristic information on the basis of the pheromone by adopting the selection method of roulette and updates local pheromone. After the heuristic information is selected, the ants map tasks in a task list to the service process in sequence on the basis of the pheromone and the heuristic information. When all the ants complete the construction of solutions, each scheduling solution is evaluated and all the pheromone is updated.

Description

The utilization ant group algorithm is optimized the method for multi-QoS grid workflow

Technical field:

The present invention relates to grid computing and intelligent algorithm two big fields, relate generally to a kind of method of using ant group algorithm to optimize the multi-QoS grid workflow.

Technical background:

Grid computing can be supported parallel on a large scale and Distributed Calculation, is considered to potential computing platform of future generation.(electrical power grid) is similar with electrical network, and computing grid makes the sharing of the various various computing resources be distributed in diverse geographic location, selection and Collaboration become possibility.Grid can satisfy science and coml computation requirement, and provides understanding to have decided to calculate the feasible method of intensive problem.

The computing application of handling in grid is commonly referred to workflow.In the environment of grid, workflow defining is: a specific task sequence of finishing complex target.In grid computing, how traffic control stream is an important problems to reach higher performance.Usually, workflow provides with the form of directed acyclic graph (DAG), and wherein node is represented individual task and oriented arrow is represented the dominance relation between task.Task dispatch need handle to Task Distribution different Distributed Calculation websites, to satisfy requirement and the computation optimization performance of user to service quality (QoS).Generally speaking, the scheduling problem in DAG be NP-completely, so workflow schedule is a very complicated problems.

The existing research major part of the workflow schedule problem in the grid concentrated on only have the qos parameter problem of (total duration).Wherein, the most frequently used method is based on the dynamic or static list scheduling algorithm of different heuristic informations, OLB algorithm for example, MET algorithm, MCT algorithm, Min-min algorithm, Max-min algorithm, Duplex algorithm, Sufferage algorithm and HEFT algorithm etc.These basic idea are to come scheduler task successively according to heuristic information.In addition, also has some Asias heuristic (meta-heuristic) algorithm, genetic algorithm (GA) for example, simulated annealing (SA) and genetic mimic anneal (GSA) etc.Generally speaking, inferior heuritic approach has more performance, but needs longer computing time.

Recently, the proposition of Open Grid Service architecture (OGSA) makes grid computing technology obtain reinforcement.OGSA is incorporated into the network service in the grid model, and becomes prevailing technology rapidly.Under new structural system, the traditional scheduler algorithm that more than provides is no longer suitable.This be because: the first, the main trend of OGSA is the various mesh services work station (GSPs) that computational resource is provided to occur.Therefore, task executions no longer is subject to the quantity of computational resource.Under new system, can finish by a series of service processes that the different operating station provides.So scheduler program just need connect each task and its corresponding service processes.The second, total duration no longer is unique qos parameter that the user is concerned about.Accordingly, other qos parameter, expense for example, reliability and fail safe etc., it is more and more important to seem in grid performance.Therefore, scheduler program need be considered the balance between the different QoS parameter, with demand that satisfies the user and the performance of optimizing whole application program.

In order to solve this more difficult scheduling problem, Yuan etc. have proposed a dynamic lattice point dispatching algorithm with consideration expense and the balance between the time [1].But this algorithm can only carry out the optimization of duration and ignore the selection of user to other parameter.Yu etc. have proposed the workflow schedule algorithm [2] of a kind of Deadline-MDP of being called.In this algorithm, the task in the workflow is divided into different branches, and each branch is endowed a sub-time limit.Use Markov decision process (MDP) to seek the scheduling that each branch has least cost and satisfies time-constrain then, but this algorithm still can only be used to solve the Cost Optimization problem of band time limit constraint.

Along with the continuous development of grid computing, the scale of workflow problem becomes bigger and bigger.And user's also can be workflow application according to the demand of self set a series of QoS constraint.Therefore, a good scheduler program should have the ability to find out optimum scheduling scheme, satisfies the qos parameter of user-defined constraints and optimization user's request.In order to solve this complicated scheduling problem, this paper has proposed a kind of method of using ant group algorithm to optimize the multi-QoS grid workflow.Ant group algorithm is a kind of inferior heuritic approach that is proposed by Dorigo.It is worked by the foraging behavior that imitates true ant, and has successfully applied in the optimization of various complex combination problems.The main advantage that ant group algorithm is used for the workflow schedule problem is that it can make full use of heuristic information based on process.

List of references:

[1]Y.Yuan，X.Li，and?Q.Wang：Time-cost?trade-off?dynamic?scheduling?algorithmfor?workflows?in?grids.Proceedings?of?the?10 ^th?International?Conference?on?ComputerSupported?Cooperative?Work?in?Design，2006。

[2]J.Yu，R.Buyya，and?C.K.Tham：Cost-based?scheduling?of?scientific?workflowapplications?on?utility?grids.Proceedings?of?the?1 ^st?International?Conference?one-Science?and?Grid?Computing(e-Science’05)，pp.140-147，2005。

Summary of the invention:

The present invention mainly concentrates on the large-scale many QoS workflow schedule problem that solves.Three basic qos parameters of main consideration in the model of setting up: reliability, time and expense.These parameters are crucial in grid application, and their characteristic also has nothing in common with each other.The user can specify the requirement to QoS when submitting workflow application: the lower bound of reliability, working life and expense budget etc.In addition, the user also may need to optimize one of them qos parameter.Therefore, the target of the method for proposition is to find out a feasible dispatching method, satisfies user-defined all qos parameters, and one of them is optimized.In order to finish this task, designed the search behavior that seven kinds of heuristic informations based on process are used to guide ant.In the algorithm of invention, use adaptive method to handle these heuristic informations, in construction solution, allow ant usually select heuristic information according to information.The concrete steps of algorithm comprise:

(1) initiation parameter.For different QoS demands, the initial value τ of pheromones ₀Provide according to following formula

Wherein, min_Reliability represents the least reliability of all service processes, and max_Reliability=100.Min_Makespan is the minimum duration of estimating of workflow, and max_Makespan is the maximum duration of estimating.These two estimated values can be by calculating each duty mapping to having in the minimum service processes of (perhaps maximum) time of implementation.Same, min_Cost and max_Cost are least cost and the costs on the higher scale that workflow is estimated.They can be by calculating each duty mapping in the service processes that has minimum (perhaps maximum) execution cost.

(2) all ants of initialization.In when beginning circulation each time, every ant all can be selected a structural grain (forward direction or back to) at random.The forward direction ant can be according to given dominance relation construction solution in work flow network.Opposite, the back will begin search from end node to ant, and all arrows is reverse.Make algorithm can search for how different separating since the scheduling strategy of two kinds of directions.In addition, every ant all will be set up a task sequence that meets precedence constraint.And task is mapped in the service processes successively based on this task sequence.Task sequence is to set up by the task of selecting at random to satisfy precedence constraint.

(3) every ant is at first selected a kind of heuristic information based on the system of selection of pheromones utilization roulette.After ant has been selected a kind of heuristic information, heuristic information A for example, the pheromones on this heuristic information will be carried out local renewal

τ _A＝(1-ρ)·τ _A+ρ·τ ₀

Wherein, ρ ∈ (0,1) is a parameter.

(4) selected a kind of heuristic information after, ant just can begin to construct separating of scheduling problem.In each step of construction solution, ant can be selected a service processes based on pheromones and heuristic information, and with first unmapped duty mapping in the task queue in this service processes.Selection is with task T _iBe mapped to S _i ^jMethod be

In this system of selection, generate a random number q ∈ [0,1], and with itself and a parameter q ₀∈ [0,1] relatively.If q≤q ₀, ant will be with T so _iBe mapped to and contain maximum τ _Ijβ ^{η ij}In the service processes of value.Otherwise,, just select S with the system of selection of utilization roulette _i ^jPossibility and τ _Ijβ ^{η ij}Size is directly proportional.τ wherein _IjAnd η _IjRepresent T respectively _iBe mapped to process s _i ^jPheromones and heuristic information, β 〉=1st, the parameter of decision pheromones relative influence with heuristic information.When ant with a task T _iBe mapped to a service processes s _i ^jAfter, carry out the renewal of local pheromones at once.The plain method for updating of local message is given by the following formula

τ _ij＝(1-ρ)·τ _ij+ρ·τ ₀

In circulation each time, every ant recycles this selection mode N time, separates thereby N duty mapping formed a complete scheduling in N the service processes.

(5) after all ants are all finished the structure of separating, each scheduling separated estimate.

(6) renewal of execution global information element.The globally optimal solution that only has high evaluation value just can carry out the renewal of global information element.Suppose that globally optimal solution is K (K ₁..., K _n), overall update strategy is given as follows

${\mathrm{\τ}}_{{\mathrm{iK}}_i}=(1-\mathrm{\ρ})\·{\mathrm{\τ}}_{{\mathrm{iK}}_i}+\mathrm{\ρ}\·K.\mathrm{score},$ i＝1，2，…，n

(7), otherwise get back to step (2) if reach termination condition then termination routine.

The method of invention has the following advantages: 1, considered three kinds of qos parameters, except the constraint of satisfying QoS, algorithm can also be optimized one of them qos parameter.2, make full use of the characteristics of ant group algorithm, designed the search that seven kinds of heuristic informations are used to guide ant, made the search efficiency to have obtained effective raising.

Description of drawings

Fig. 1 workflow application process schematic diagram

The process schematic diagram of a global solution of Fig. 2 ant structure

Fig. 3 algorithm overall flow figure

Embodiment:

Below in conjunction with accompanying drawing, further the method to invention is described.

Problem at science or a lot of computation-intensives of coml can be classified as the workflow problem.The process that workflow is used as shown in Figure 1.The user at first can provide the abstract language that workflow is used in the abstractdesription aspect.Grid system just needs to select and application component is set to form abstract workflow, the order that the formulation task is carried out in abstract workflow then.In OSGA, the task in the workflow is to be carried out by the network service processes that work station provides.The different processes that provided by the different operating station can be used to carry out same task, and have different qos parameters.So, the task of the 3rd aspect be exactly with the duty mapping in the abstract workflow in service processes, to generate concrete workflow.For example in Fig. 1, each the task T in abstract workflow _iAll corresponding a series of service processes $S_i=\{s_i^1,s_i^2,\·\·\·s_i^{m_i}\}.$ s _i ¹, s _i ²..., s _i ^MiAt rectangle S _iIn with circle expression.m _iBe task T _iThe sum of pairing service processes.The service processes of representing with gray circles is the process of last this task of execution of selecting, has so just constituted concrete workflow.At last, at application, different work stations may be carried out same program with different structures.In this model, abstract workflow is mapped as the step that concrete workflow is a most critical in the grid application.Target of the present invention is exactly to propose an effective algorithm to generate concrete workflow for large-scale workflow application problem.

Intuitively, the effect of workflow schedule algorithm be with all duty mapping in the abstract workflow to service processes, generating optimum concrete workflow, and satisfy the performance of user-defined QoS constraint and maximization user's request.At the DAG of abstract workflow, (V A) in the model, makes that n is the quantity of task in the workflow to G=.The set V={T of node ₁, T ₂..., T _nTask in the corresponding workflow.The set A of arrow is represented the order of priority relation between the task.Article one, arrow is with (T _i, T _j) expression, wherein T _iBe called T _jFather's task, and T _jBe called T _iThe subtask.Under normal circumstances, a subtask can only could begin to carry out after its all father's tasks are all finished.T _iThe set of father's task with Pred (T _i) expression, and T _iThe set of subtask then use Succ (T _i) expression.

Each task T _i(1≤i≤n) has an execution scope $S_i=\{s_i^1,s_i^2,\·\·\·s_i^{m_i}\},$ S wherein _i ^j(1≤j≤m _i) service processes providing by work station of expression, m _iExpression T _iThe quantity of available service processes.The attribute of a service processes can be represented (s with one group of four variable _i ^j.g, s _i ^j.r, s _i ^j.t, s _i ^j.c).Wherein, s _i ^j.g represent s _i ^jAffiliated work station is s _i ^j.g.s _i ^j.r, s _i ^j.t and s _i ^j.c represent s respectively _i ^jReliability, time of implementation and expense.

Precedence constraint has been included in the DAG model, and it has specified a task execution order that feasible solution must satisfy.

In addition, in Work flow model, also there are some user-defined QoS constraints.Since considered three kinds of qos parameters in this article, therefore to there being three types QoS to retrain:

(1) reliability constraint:

The reliability of the concrete workflow that generates can not be less than a user-defined variables L imitedReliability.In other words, a given scheduling scheme K (K ₁..., K _n) (K _iExpression T _iBy s _i ^KiCarry out), if K.reliability is the reliability of this scheduling, the condition that scheduling K satisfies reliability constraint is

$K.\mathrm{reliability}={\mathrm{<figure-callout\; id="1"\; label="min"\; filenames="S2008100264875E00031.png"\; state="\{\{state\}\}">min</figure-callout>}}_{1\&le;i\&le;n}{s}_i^{k_i}.r\&GreaterEqual;\mathrm{LimitedReliability}$

(2) total duration constraint:

Total time of implementation of workflow can not be greater than a user-defined variables D eadline.In other words, if K.makespan is total time of implementation of K, K should satisfy so

K.makespan≤Deadline

(3) expense restriction:

For given scheduling scheme K (K ₁..., K _n), the total cost of K (K.cost) can not be greater than a user-defined variable Budget.Just

$K.\cos t=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{s}_i^{K_i}.c\&le;\mathrm{Budget}$

Except constraints, the user also can define the qos parameter that needs optimization.

(1) reliability optimization:

The user tends to the reliability of optimization application.In this case, the user tends to be provided with the constraint of duration and expense, and the target of dispatching algorithm then is to seek a scheduling scheme K who satisfies all constraintss, and the value of maximization K.reliability.

(2) duration is optimized:

The user tends to the duration of optimization application.In this case, the user tends to be provided with the constraint of reliability and expense, and the target of dispatching algorithm then is to seek a scheduling scheme K who satisfies all constraintss, and minimizes the value of K.makespan.

(3) Cost Optimization:

The user tends to the expense of optimization application.In this case, the user tends to be provided with the constraint of reliability and duration, and the target of dispatching algorithm then is to seek a scheduling scheme K who satisfies all constraintss, and minimizes the value of K.cost.

Below describe the ant group algorithm that is used to solve this scheduling problem in detail:

1. the definition of pheromones and heuristic information

For ant group algorithm, pheromones and heuristic information are most important factors.Because scheduling problem is therefore to establish task T with all duty mapping in the abstract workflow in service processes _iBe mapped to process s _i ^jPheromones be τ _Ij, heuristic information is η _Ij

When algorithm began, establishing all pheromones was an initial value τ ₀, just

τ _ij＝τ ₀(1≤i≤n，1≤j≤m _i)

In addition, owing in the model of considering, have a plurality of qos parameters, so designed seven kinds of heuristic informations in the algorithm.

(1) heuristic information A: reliability greediness (RG)

The RG heuristic information makes artificial ant be more prone to select to have the more process of high reliability.That suppose the ant use is RG heuristic information, T so _iBe mapped to s _i ^jHeuristic information (use RG _IjExpression) is

${\mathrm{\&eta;}}_{\mathrm{ij}}={\mathrm{RG}}_{\mathrm{ij}}=\frac{s_i^j.r-\min \_\mathrm{reliabilit}{y}_i+1}{\max \_\mathrm{reliabilit}{y}_i-\min \_\mathrm{reliabilit}{y}_i+1}$

Wherein, $\min \_\mathrm{reliabilit}{y}_i={\mathrm{<figure-callout\; id="1"\; label="min"\; filenames="S2008100264875E00031.png"\; state="\{\{state\}\}">min</figure-callout>}}_{1\&le;j{\&le;m}_i}\{s_i^j.r\},$ $\max \_\mathrm{reliabilit}{y}_i={\mathrm{<figure-callout\; id="1"\; label="max"\; filenames="S2008100264875E00031.png"\; state="\{\{state\}\}">max</figure-callout>}}_{1\&le;j{\&le;m}_i}\{s_i^j.r\}.$ According to formula, one has more that the process of high reliability will have higher heuristic information.Simultaneously, this formula has also guaranteed η _Ij∈ (0,1].

(2) heuristic information B: time greediness (TG)

The TG heuristic information makes artificial ant be more prone to select shorter process of time of implementation.That suppose the ant use is TG heuristic information, T so _iBe mapped to s _i ^jHeuristic information (use TG _IjExpression) is

${\mathrm{\&eta;}}_{\mathrm{ij}}={\mathrm{TG}}_{\mathrm{ij}}=\frac{\max \_\mathrm{tim}{e}_i-s_i^j.t+1}{\max \_\mathrm{tim}{e}_i-\min \_\mathrm{tim}{e}_i+1}$

Wherein, $\min \_\mathrm{tim}{e}_i={\mathrm{<figure-callout\; id="1"\; label="min"\; filenames="S2008100264875E00031.png"\; state="\{\{state\}\}">min</figure-callout>}}_{1\&le;j{\&le;m}_i}\{s_i^j.t\},$ $\max \_{\mathrm{time}}_i={\mathrm{<figure-callout\; id="1"\; label="max"\; filenames="S2008100264875E00031.png"\; state="\{\{state\}\}">max</figure-callout>}}_{1\&le;j{\&le;m}_i}\{s_i^j.t\}.$ According to formula, a shorter process of time of implementation will have higher heuristic information, and η _Ij∈ (0,1].

(3) heuristic information C: expense greediness (CG)

The CG heuristic information makes artificial ant be more prone to select required expense process still less.That suppose the ant use is CG heuristic information, T so _iBe mapped to s _i ^jHeuristic information (use CG _IjExpression) is

${\mathrm{\&eta;}}_{\mathrm{ij}}={\mathrm{CG}}_{\mathrm{ij}}=\frac{\max \_\cos t_i-s_i^j.t+1}{\max \_\cos t_i-\min \_\cos t_i+1}$

Wherein, $\min \_\cos t_i={\mathrm{<figure-callout\; id="1"\; label="min"\; filenames="S2008100264875E00031.png"\; state="\{\{state\}\}">min</figure-callout>}}_{1\&le;j\&le;m_i}\{s_i^j.t\},$ $\max \_{\cos t}_i={\mathrm{<figure-callout\; id="1"\; label="max"\; filenames="S2008100264875E00031.png"\; state="\{\{state\}\}">max</figure-callout>}}_{1\&le;j\&le;m_i}\{s_i^j.t\}.$ According to formula, required expense process still less will have higher heuristic information, and η _Ij∈ (0,1].

(4) heuristic information D: the time limit of suggestion (SD)

Between qos parameter, always exist and accept or reject balance.For example, short process of time of implementation may need higher expense and reliability lower.Consider the such balance and the constraint in time limit, the SD heuristic information makes artificial ant be more prone to select the service processes of just on time finishing.In order to reach this target,, give the time limit of a suggestion to each task in abstract workflow based on the user-defined time limit.

In order to calculate the SD heuristic information of each task, at first need to calculate the earliest start time of each task and back to earliest start time.

Task T _iEarliest start time (EST _i): with each task T _iBe mapped in the shortest service processes of time of implementation EST _iEqual task T under this mapping _iTime started.In addition, the workflow under this mapping method is carried out minimum total duration that total duration can be counted as estimating, and represents with min_Makespan.

Task T _iBack to earliest start time (BEST _i): by with the start node among the DAG as end node, with end node node to start with, and with the direction counter-rotating of all arrows, just DAG can be converted to one after to network.For each task T _i, the EST in the back in network _iValue is exactly BEST _i

Based on above two times, just can calculation task T _iThe on average the shortest time of implementation

$\mathrm{avg}\_\min \_\mathrm{tim}{e}_i=\frac{\left({\min }_{{\&ForAll;T}_j\&Element;\mathrm{succ}\left(T_i\right)}\right\{{\mathrm{EST}}_j\}-{\mathrm{EST}}_i)+({\min }_{{\&ForAll;T}_j\&Element;\mathrm{pred}\left(T_i\right)}{\mathrm{BEST}}_j-{\mathrm{BEST}}_i)}{2}$

According to avg_min__time _iCalculate SD _iMethod be

That suppose the ant use is SD heuristic information, T so _iBe mapped to s _i ^jHeuristic information be

${\mathrm{\&eta;}}_{\mathrm{ij}}=\frac{\max \left\{\right|\max \_\mathrm{tim}{e}_i-{\mathrm{SD}}_i|,|{\mathrm{SD}}_i-\min \_\mathrm{tim}{e}_i\left|\right\}-(s_i^j.t-{\mathrm{SD}}_i)+1}{\max \left\{\right|\max \_\mathrm{tim}{e}_i-{\mathrm{SD}}_i|,|{\mathrm{SD}}_i-\min \_\mathrm{tim}{e}_i\left|\right\}+1}$

According to formula, a time of implementation is more near SD _iProcess will have higher heuristic information, and η _Ij∈ (0,1].

(5) heuristic information E: the budget of suggestion (SB)

Similar to the SD heuristic information, the SB heuristic information makes ant be more prone to select just to reach the service processes of estimated cost.In order to reach this target,, give the budget of a suggestion to each task in abstract workflow based on user-defined master budget.

By with all duty mapping in the service processes of minimum charge, obtain the least cost min_Cost of whole workflow.Just, $\min \_\mathrm{Cost}=\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}\min \_{\cos t}_j.$ Task T _iSuggestion budget SB _iFollowing calculating

${\mathrm{SB}}_i=\min \_\cos t_i\&CenterDot;\frac{\mathrm{Budget}}{\min \_\mathrm{Cost}}$

What suppose the ant use is the SB heuristic information, so with T _iBe mapped to s _i ^jHeuristic information be

${\mathrm{\&eta;}}_{\mathrm{ij}}=\frac{\max \left\{\right|\max \_{\cos t}_i-{\mathrm{SB}}_i|,|{\mathrm{SB}}_i-\min \_{\cos t}_i\left|\right\}-(s_i^j.c-{\mathrm{SB}}_i)+1}{\max \left\{\right|\max \_{\cos t}_i-{\mathrm{SB}}_i|,|{\mathrm{SB}}_i-\min \_{\cos t}_i\left|\right\}+1}$

According to formula, expense is more near SB _iService processes will have higher heuristic information, and η _Ij∈ (0,1].

(6) heuristic information F: time/expense (TC)

The TC heuristic information is taken all factors into consideration the time and the cost element of a service processes.It combines TG and two kinds of heuristic informations of CG.What suppose the ant use is the TC heuristic information, so task T _iBe mapped to s _i ^jHeuristic information be

${\mathrm{\&eta;}}_{\mathrm{ij}}=\frac{1}{2}\&CenterDot;({\mathrm{TG}}_{\mathrm{ij}}+{\mathrm{CG}}_{\mathrm{ij}})$

According to formula, the service processes with shorter time of implementation and lower expense will have higher heuristic information, and η _Ij∈ (0,1].

(7) heuristic information G: overall performance (OP)

In the OP heuristic information, taken all factors into consideration the effect (comprising reliability, time and expense) of all qos parameters.It combines TG, CG and RG heuristic information.What suppose the ant use is the OP heuristic information, so task T _iBe mapped to s _i ^jHeuristic information be

${\mathrm{\&eta;}}_{\mathrm{ij}}=\frac{1}{3}\&CenterDot;({\mathrm{TG}}_{\mathrm{ij}}+{\mathrm{CG}}_{\mathrm{ij}}+{\mathrm{RG}}_{\mathrm{ij}})$

According to formula, have the shorter time of implementation, lower expense and more the service processes of high reliability will have higher heuristic information, and η _Ij∈ (0,1].

Based on different user's requests, algorithm will use different heuristic informations.1) if target is to optimize reliability, algorithm will use seven kinds of all heuristic informations so.Wherein, RG and OP heuristic information are used to seek the service processes of high reliability, and other heuristic information then is used for guarantee fee with satisfying the QoS constraint with the duration.2) if target is to optimize the duration, so just only use heuristic information TG, CG, SB and TC.TG and TC heuristic information are used to seek the service processes of the shortest time of implementation, and CG and SB then are used to search for the service processes that satisfies expense restriction.In this case, the constraints of reliability can reach by the service processes of not selecting to be lower than reliability constraint, does not therefore need RG and OP heuristic information.Because the target of algorithm is to minimize the duration, so also do not need the SD heuristic information.3) if target is the optimization expense, based on similar reason, required heuristic information is TG, CG, SD and TC.

Use adaptive method in algorithm, to manage this seven kinds of heuristic informations.Before artificial ant begins construction solution at every turn, at first can select the used heuristic information of this structure based on pheromones.Wherein, the pheromones τ of corresponding heuristic information A _AExpression, the pheromones τ of corresponding heuristic information B _BExpression, by that analogy.When algorithm began, the pheromones on all heuristic informations that need will be set to initial value τ ₀, the pheromones of other no heuristic information then is made as 0.

2. the structure of separating

In each circulation of algorithm, the ant that all can to use one group of number be M comes construction problem to separate.The constructor of separating can be divided into following two steps.

(1) initialization of ant

In when beginning circulation each time, every ant all can be selected a structural grain (forward direction or back to) at random.The forward direction ant can be according to given dominance relation construction solution in work flow network.Opposite, the back will begin search from the end node of DAG to ant, and all arrows is reverse, as calculating the back to the earliest start time.Make algorithm can search for how different separating since the scheduling strategy of two kinds of directions.

In addition, every ant all will be set up a task sequence that meets precedence constraint.And task is mapped in the service processes successively based on this task sequence.For convenience's sake, the task sequence with the forward direction ant is expressed as (T _Start, T _Q1, T _Q2..., T _Qn, T _End), the back is expressed as (T to the task sequence of ant _End, T _Q1, T _Q2..., T _Qn, T _Start).Task sequence is to set up by the task of selecting at random to satisfy precedence constraint.

(2) the structure scheduling is separated

In this step, M ant M of construction problem respectively separated.The process of a global solution of ant structure as shown in Figure 2.When beginning, each ant can use the system of selection of roulette to select a kind of heuristic information based on pheromones.Just, select probability and the pheromones τ of heuristic information A _AShared ratio is directly proportional.

After having selected a kind of heuristic information, ant just begins to construct having separated of scheduling problem.In each step of construction solution, ant can be selected a service processes based on pheromones and heuristic information, and with first unmapped duty mapping in the task queue in this service processes.Selection is with task T _iBe mapped to S _i ^jMethod be

In this system of selection, generate a random number q ∈ [0,1], and with itself and a parameter q ₀∈ [0,1] relatively.If q≤q ₀, ant will be with T so _iBe mapped to and contain maximum τ _Ijβ ^{η ij}In the service processes of value.Otherwise,, just select S with the system of selection of utilization roulette _i ^jPossibility and τ _Ijβ ^{η ij}Size is directly proportional.Wherein, β 〉=1st, the parameter of the relative influence of decision pheromones with heuristic information.

In circulation each time, every ant recycles this selection mode N time, separates thereby N duty mapping formed a complete scheduling in N the service processes.

3. the processing of pheromones

(1) pheromones initialization

In this article, for different QoS demands, the initial value τ of pheromones ₀Provide according to following formula

Wherein, min_Reliability represents the least reliability of all service processes, and max_Reliability=100.Min_Makespan is the minimum duration of estimating of workflow, and max_Makespan is the maximum duration of estimating.These two estimated values can be by calculating each duty mapping to having in the minimum service processes of (perhaps maximum) time of implementation.Same, min_Cost and max_Cost are least cost and the costs on the higher scale that workflow is estimated.They can be by calculating each duty mapping in the service processes that has minimum (perhaps maximum) execution cost.

(2) local message is plain upgrades

The invention method in, when ant with a task T _iBe mapped to a service processes s _i ^jAfter, carry out the renewal of local pheromones at once.The plain method for updating of local message is given by the following formula

τ _ij＝(1-ρ)·τ _ij+ρ·τ ₀

Wherein, ρ ∈ (0,1) is a parameter.Because τ ₀Also be the minimum of pheromones simultaneously, so the plain effect of upgrading of local message is to reduce τ _IjValue to increase the diversity of algorithm.

Pheromones on the heuristic information equally also will be carried out the renewal of local message element.After ant has been selected a kind of heuristic information, heuristic information A for example, the pheromones on this heuristic information will change according to following formula

τ _A＝(1-ρ)·τ _A+ρ·τ ₀

(3) global information is plain upgrades

All constructed the renewal of carrying out the global information element after separating when all ants.Algorithm at first can be separated in the present age circulation all and compare.The quality that K is separated in scheduling can calculate according to following three formulas.For the reliability optimization problem, the evaluation of estimate of K is

Wherein, the evaluation of estimate of K is made up of two parts: the firstth, satisfy the evaluation of estimate that QoS retrains, and the secondth, the evaluation of estimate of the quality of the qos parameter that need optimize.The interval of each part evaluation be (0,1], thus the interval of K.score be (0,2].If K has satisfied all QoS constraints, the evaluation of estimate of corresponding QoS constraint will be made as 1, and the evaluation of estimate of the qos parameter that the user need optimize is then set according to the reliability of K.On the other hand, if K can not satisfy all QoS constraints, the evaluation of estimate of QoS constraint will be set according to the degree that satisfies constraint, and the evaluation of estimate of the qos parameter that the user need optimize then can be set to minimum value.

Similarly, in duration optimization, the evaluation of estimate of K is

For Cost Optimization, the evaluation of estimate of K is

The globally optimal solution that only has high evaluation value just can carry out the renewal of global information element.Suppose that globally optimal solution is K (K ₁..., K _n), overall update strategy is given as follows

${\mathrm{\&tau;}}_{{\mathrm{iK}}_i}=(1-\mathrm{\&rho;})\&CenterDot;{\mathrm{\&tau;}}_{{\mathrm{iK}}_i}+\mathrm{\&rho;}\&CenterDot;K.\mathrm{score},$ i＝1，2，…，n

The flow chart of whole algorithm as shown in Figure 3.

Algorithm according to user-defined QoS constraint and the extensive grid work flow of demand optimization is very rare, in existing document, has only the Deadline-MDP algorithm of proposition such as Yu can be used for finding the solution of similar problem.Therefore, method and the Deadline-MDP algorithm with invention compares.Be noted that the Deadline-MDP algorithm can only be used to solve the Cost Optimization problem of band time limit constraint.10 examples that the utilization grid work flow is used come these two kinds of methods are tested.Wherein, preceding 3 examples are the workflow problems in the reality, comprise e-Economic application problem, Neuscience application problem and e-protein workflow problem.Remaining example then generates according to the PSPLIB storehouse.Because Deadline-MDP is a kind of deterministic algorithm, so it can only provide one of problem to separate.The ant group algorithm that proposes then is a kind of random algorithm of tape guide.In order to compare all sidedly, the method independent operating of inventing is obtained 100 results for 100 times.In 10 all problems, the mean value that ant group algorithm obtains all is better than the result that Deadline-MDP obtains.In addition, in or (have 30 above tasks) in the extensive problem, even the poorest result that ant group algorithm obtains also is better than the result of Deadline-MDP.Generally speaking, ant group algorithm can reduce the expense of workflow 10-20%, and this method that has proved invention is effective.

Claims (4)

1, a kind of method of using ant group algorithm to optimize the multi-QoS grid workflow is characterized in that this method may further comprise the steps:

(1) initiation parameter.For different QoS demands, the initial value τ of pheromones ₀Provide according to following formula

Wherein, min_Reliability represents the least reliability of all service processes, and max_Reliability=100.Min_Makespan is the minimum duration of estimating of workflow, and max_Makespan is the maximum duration of estimating.These two estimated values can be by calculating each duty mapping to having in the minimum service processes of (perhaps maximum) time of implementation.Same, min_Cost and max_Cost are least cost and the costs on the higher scale that workflow is estimated.They can be by calculating each duty mapping in the service processes that has minimum (perhaps maximum) execution cost.

(2) all ants of initialization.In when beginning circulation each time, every ant all can be selected a structural grain (forward direction or back to) at random.The forward direction ant can be according to given dominance relation construction solution in work flow network.Opposite, the back will begin search from end node to ant, and all arrows is reverse.In addition, every ant all will be set up a task sequence that meets precedence constraint.And task is mapped in the service processes successively based on this task sequence.Task sequence is to set up by the task of selecting at random to satisfy precedence constraint.

(3) every ant is at first selected a kind of heuristic information based on the system of selection of pheromones utilization roulette.After ant has been selected a kind of heuristic information, heuristic information A for example, the pheromones on this heuristic information will be carried out local renewal

τ _A＝(1-ρ)·τ _A+ρ·τ ₀

Wherein, ρ ∈ (0,1) is a parameter.

(4) selected a kind of heuristic information after, ant just can begin to construct separating of scheduling problem.In each step of construction solution, ant can be selected a service processes based on pheromones and heuristic information, and with first unmapped duty mapping in the task queue in this service processes.Selection is with task T _iBe mapped to S _i ^jMethod be

In this system of selection, generate a random number q ∈ [0,1], and with itself and a parameter q ₀∈ [0,1] relatively.If q≤q ₀, ant will be with T so _iBe mapped to and contain maximum τ _Ijβ ^{η ij}In the service processes of value.Otherwise,, just select S with the system of selection of utilization roulette _i ^jPossibility and τ _Ijβ ^{η ij}Size is directly proportional.τ wherein _IjAnd η _IjRepresent T respectively _iBe mapped to process s _i ^jPheromones and heuristic information, β 〉=1st, the parameter of decision pheromones relative influence with heuristic information.When ant with a task T _iBe mapped to a service processes s _i ^jAfter, carry out the renewal of local pheromones at once.The plain method for updating of local message is given by the following formula

τ _ij＝(1-ρ)·τ _ij+ρ·τ ₀

In circulation each time, every ant recycles this selection mode N time, separates thereby N duty mapping formed a complete scheduling in N the service processes.

(5) after all ants are all finished the structure of separating, each scheduling separated estimate.

(6) renewal of execution global information element.The globally optimal solution that only has high evaluation value just can carry out the renewal of global information element.Suppose that globally optimal solution is K (K ₁..., K _n), overall update strategy is given as follows

${\mathrm{\&tau;}}_{{\mathrm{iK}}_i}=(1-\mathrm{\&rho;})\&CenterDot;{\mathrm{\&tau;}}_{{\mathrm{iK}}_i}+\mathrm{\&rho;}\&CenterDot;K.\mathrm{score},$ i＝1，2，…，n

(7), otherwise get back to step (2) if reach termination condition then termination routine.

2, based on the described a kind of method of using ant group algorithm to optimize the multi-QoS grid workflow of claim 1, it is characterized in that this method has been considered three basic qos parameters: reliability, time and expense.The user can provide required QoS constraint, and the qos parameter of specifying needs to optimize.

3, based on the described a kind of method of using ant group algorithm to optimize the multi-QoS grid workflow of claim 1, it is characterized in that, designed the search behavior that seven kinds of heuristic informations based on process are used to guide ant.Ant at first needs to select a kind of heuristic information to guide this construction process according to pheromones when construction solution.

4, based on the described a kind of method of using ant group algorithm to optimize the multi-QoS grid workflow of claim 1, it is characterized in that the evaluation of estimate K.score that separates K can calculate according to following formula.For the reliability optimization problem, the evaluation of estimate of K is

For the duration optimization problem, the evaluation of estimate of K is

For the Cost Optimization problem, the evaluation of estimate of K is

In these three formulas, K.cost is the total cost of K, the K reliability that K.reliability is, and K.makespan is total time of implementation of K.Budget is user-defined costs on the higher scale constraint, and LimitedReliability is user-defined least reliability constraint, and Deadline is the user-defined time limit constraint of finishing.

Images